I work on data-driven reachability analysis for verifying robot controllers. Instead of relying on a model of the controller, I use its own executions to estimate the states from which it reaches its goal without failing, and how likely it is to do so when the dynamics are noisy. These estimates can be calibrated to carry statistical guarantees. Because they only need data, they work for black-box controllers such as reinforcement learning or behavior cloning policies, whose internals can't be analyzed directly.
This builds on my earlier work in the lab on planning and control for robots with complex dynamics, and on estimating controllers' regions of attraction from data.
PhD student in Prof. Kostas Bekris's PRACSYS Lab, Rutgers University
Keeps a robot on its planned trajectory despite noisy observations by jointly estimating its recent trajectory and adapting its controls. Both update online at 10 Hz or more.
I implemented the STEAP (Simultaneous Trajectory Estimation and Planning) baseline.
A survey of 50 years of methods for generating robot motion, from those built on explicit models to those that learn implicit ones, and the opportunities to combine them.
I worked on the sections on explicit motion planning and control, and on reinforcement learning and behavior cloning.
Executes planned trajectories safely with only a rough dynamics model tuned from data, combining kinodynamic replanning, feedback control, and a safety mechanism, for mobile robots on surfaces with unknown friction.
I co-developed the method, proposed and implemented its safety mechanism (contingency checking), and ran the physical experiments on the MuSHR platform.
Speeds up planning for robots with complex dynamics by connecting a learned controller through a precomputed roadmap. Tested on vehicles driving over uneven terrain and on a quadrotor.
I trained the quadrotor RL policy, and implemented and ran the RL-only baseline without the planner.
Estimates from data where a robot controller succeeds, known as its region of attraction, by analyzing its dynamics in a learned latent space. It scales to a 67-dimensional humanoid and a 96-dimensional three-fingered manipulator.
I proposed and implemented the label-based contrastive training variant.
Classifies land cover in multifrequency radar (PolSAR) images using autoencoder features and superpixels.
NSF-NRT SOCRATES Fellow (2024 – 2026)
Thesis: Safety and Efficiency for High-dimensional Robots: Integrating Learning, Planning, and Control
Outstanding Project and Outstanding Publication Awards from Rutgers CS for MORALS (2024)
Expected to graduate in May 2029, and open to internships and collaborations.
